Practical Facial Landmarks Detector

My unofficial implementation of PFLD paper "Practical Facial Landmarks Detector" using Pytorch for a real time landmarks detection and head pose estimation.

Demo

How to Install

 $ pip3 install -r requirements.txt
 
 # Note that it can be run on lower versions of Pytorch so replace the versions with yours

install opencv & dnn from source (optional)

Both opencv dnn & haar cascade are used for face detection, if you want to use haar cascade you can skip this part.

sudo apt update && sudo apt install -y cmake g++ wget unzip
wget -O opencv.zip https://github.com/opencv/opencv/archive/master.zip
wget -O opencv_contrib.zip https://github.com/opencv/opencv_contrib/archive/master.zip
unzip opencv.zip
unzip opencv_contrib.zip
mkdir -p build && cd build
cmake -DOPENCV_EXTRA_MODULES_PATH=../opencv_contrib-master/modules ../opencv-master
cmake --build .

if you have any problems, refere to Install opencv with dnn from source

Run Camera Demo

Live camera demo

$ python3 camera_demo.py

# add '--head_pose' option to visualize head pose directions 
$ python3 camera_demo.py --head_pose

# add '--haar' option if you want to use Haar cascade detector instead of dnn opencv face detector
$ python3 camera_demo.py --haar

WFLW Dataset

Wider Facial Landmarks in-the-wild (WFLW) contains 10000 faces (7500 for training and 2500 for testing) with 98 fully manual annotated landmarks.

Download the dataset & place it in 'data/WFLW' folder path

WFLW Training and Testing Images Google Drive
WFLW Face Annotations Download

Prepare the dataset

Dataset augumentation & preparation
(Only apply one of the 2 options) for data augumentation

$ python3 generate_dataset.py

# another option to augument dataset from polarisZhao/PFLD-pytorch repo 
$ cd data
$ python3 SetPreparation.py

Visualize dataset

Visualize dataset examples with annotated landmarks & head pose

# add '--mode' option to determine the dataset to visualize
$ python3 visualization.py

Visualize euler angles

$ python3 euler_angles.py

Tensorboard

Take a wide look on dataset examples using tensorboard

$ python3 visualization.py --tensorboard
$ tensorboard --logdir checkpoint/tensorboard

Testing on WFLW test dataset

$ python3 test.py

Training

Train on augumented WFLW dataset

$ python3 train.py

Folder structure

├── model						# model's implementation
├── data						# data folder contains WFLW dataset & generated dataset
	├── WFLW/					# extract WFLW images & annotations inside that folder
    	├── WFLW_annotations/
        ├── WFLW_images/
    ├── train					# generated train dataset
    ├── test					# generated test dataset

Refrences

MobileNet

3D-2D correspondences rotation:

Other PFLD Implementations

Survey

https://www.readcube.com/articles/10.1186%2Fs13640-018-0324-4

weak prespective projection

AmrElsersy/PFLD-Pytorch-Landmarks